Gaussian or Plane? Both: Semantic-Driven Voxel Representation for LiDAR�Inertial Odometry
Haiyang Wu, George Vosselman, Ville Lehtola
AI summary
Problem
Existing LiDAR-inertial odometry systems rely on a single geometric or probabilistic map representation, limiting their adaptability and accuracy across mixed structured and unstructured scenes.
Approach
An online semantic segmentation network classifies each voxel as planar or nonplanar, enabling the system to apply tailored geometric or probabilistic residual models during scan matching within an extended Kalman filter.
Key results
- Semantic-driven hybrid voxel representation dynamically classifies voxels as planar or Gaussian based on semantic labels.
- Scale-consistent scan matching fuses geometric and probabilistic residuals under a unified cost function.
- Demonstrates superior trajectory accuracy over state-of-the-art methods across UGV, UAV, and handheld platforms.
- Maintains real-time performance while improving localization robustness in diverse environments.
Why it matters
Enables more robust and accurate autonomous navigation for robots and drones operating in complex, mixed-structure environments without requiring scene-specific tuning.
Abstract
Accurate LiDAR-inertial odometry (LIO) highly de- pends on the geometric fidelity of the underlying environment rep- resentation. We explore the new and interesting research direction of integrating semantic segmentation models into metric odometry algorithms to enrich their representational capacity. Specifically, this letter proposes a semantic-driven hybrid voxel representa- tion in which an off-the-shelf 3D segmentation network assigns every voxel to either a planar or nonplanar class, using planar and Gaussian representations, respectively. Consequently, a hybrid scan matching strategy is presented using class-specific residual models that are tailored to the distinct error statistics of each surface category. The scan matcher is embedded within an Iterated Extended Kalman Filter (IEKF) for odometry and mapping. We evaluate our method on diverse platforms and environments, and show improved localization accuracy across various indoor and outdoor scenarios, while maintaining real-time performance.